ADVANCEMENTS IN AI

ADVANCEMENTS IN AI

Congressional Record, Volume 171 Issue 25 (Thursday, February 6, 2025) [Congressional Record Volume 171, Number 25 (Thursday, February 6, 2025)] [House] [Pages H547-H550] From the Congressional Record Online through the Government Publishing Office [ www.gpo.gov ] ADVANCEMENTS IN AI The SPEAKER pro tempore. Under the Speaker's announced policy of January 3, 2025, the Chair recognizes the gentleman from California (Mr. Kiley) for 30 minutes. Mr. KILEY of California. Mr. Speaker, I wanted to say a few words this evening about the rapid advancements that we are witnessing when it comes to the capability of artificial intelligence models. I don't refer just to some of the narrow applications that folks are familiar with, that they might take advantage of in their work, that they might fear eventually replacing them in work. These are sort of the narrow conceptions that exist in public discourse. The broader situation here involves the explicit goal of the leading labs to create what is known as artificial general intelligence, which is incredibly capable models that exceed human capability across essentially any domain. I actually find it stunning that some of the advancements that we have seen lately have gone essentially unnoticed, unreported on by the media. What we have seen just in the last few weeks, I think on several occasions, there should have been front-page stories about the dramatic advances that have been made, given the capacity these advances hold for transforming our lives, economy, society, and much else. What I wanted to do today is present just a few basic thoughts and pieces of information on the development and innovation that has been taking place. I am someone who has no technical expertise in this area at all. I just try to follow it closely because I believe the changes that will be upon us soon are so profound. To give you just kind of a flavor for the scale of change that I am talking about, here are a few quotes from leaders in the field: Sundar Pichai, the CEO of Google, said: ``AI is probably the most important thing humanity has ever worked on. I think of it as something more profound than electricity or fire.'' Demis Hassabis, the founder of DeepMind, echoed these sentiments. He said that AI should not be thought of as just another technology. He said that it is more epoch-defining than even the internet or mobile, more like electricity or fire. Sam Altman, who, of course, is the head of OpenAI, said: ``With these new abilities, we can have shared prosperity to a degree that seems unimaginable today. In the future, everyone's lives can be better than anyone's life is now. . . . Eventually, we can each have a personal AI team full of virtual experts in different areas working together to create almost anything we can imagine.'' Along those lines, Elon Musk, who, in addition to his other ventures, is the founder of xAI, said: ``AI will ultimately render money meaningless.'' Why is that? He believes the capabilities will essentially allow any person access to basically any good that they desire. Dario Amodei, who is the CEO at another lab, Anthropic, says that it is his guess that powerful AI could accelerate the rate of scientific discoveries by a factor of 10, giving us the next 50 to 100 years of biological progress in just the next 5 to 10 years. We are already seeing incredible applications such as AlphaFold from DeepMind, which Demis Hassabis recently won the Nobel Prize for, which has predicted the 3D structures of over 200 million proteins, essentially solving the protein folding problem. At the recent unveiling of the Stargate initiative, Mr. Altman was joined by Larry Ellison, who also spoke about the potential to cure cancer and essentially any other disease. This might all sound pretty vague and certainly very optimistic. Maybe it sounds like hype, so I want to present a couple of charts that really clearly make this point, not only about the ultimate potential of AI models but about the rapid acceleration that we are right now currently in the midst of. {time} 1845 This chart is one metric of basically how smart a model is. It is called the GPQA diamond. If you look at the chart here, this axis is how well it does on the test, the model. [[Page H548]] This axis is when the model was released. The different models over the time are the ``Xs'' there. The chart doesn't actually go back that far. It just goes back to July of 2023. We are looking at basically 35 percent for one of the models then. You see that the line continues to go up steadily, a little bit at a time. Now, just in the last few weeks, you have seen this truly exponential growth rate. Those are both OpenAI models that are listed there; o1 pro and o3. I guess the o2 there was some trademark issues, so they jumped right from 1 to 3. That shows you that we are on a rapidly accelerating curve. Now, the second chart I think makes this point even more clearly. One of the problems with the benchmarks that are being used to judge the capability of models is they get saturated very quickly. The models are becoming so much smarter that they ace the exam, and so the exam isn't good for anything anymore. There was this effort to come up with what they called humanity's last exam, saying once they can solve this, then they have solved just about anything. This was unveiled, and the model's GPT-4o, which is actually the model that most people use--if you just go to ChatGPT, that is the one that it sort of defaults to--it didn't so well, 3.3 percent. However, we are now up to the latest model that OpenAI has released, 26.6 percent. I included this tweet from--I think this actually is a reporter who said: When I wrote about humanity's last exam, the leading AI model got an 8.3 percent score. Five models now surpassed that, and the best model is 26.6 percent. That was 10 days ago. In 10 days, we have gone from 8.3 percent to 26.6 percent on humanity's last exam. By the way, you will see that DeepSeek here at a respectable 9.4 percent. This is kind of the exception where you have seen a lot of media coverage of an advancement in AI. The coverage really focused on the geopolitical implications of having this breakthrough come from a Chinese company. Of course, that is a very, very serious and concerning topic. Kind of lost in the coverage was perhaps the more important point of what it means for all of us, for humanity collectively, when we see this sort of exponential growth in the capability of these AI models. What is essentially driving this very recent trend of exponential growth is the advent of unlocking a new scaling law around what is called test time compute. For a while, the way that these models were getting more capable was by scaling up the compute, the computation, that went into training them. You trained the model, and it became more and more powerful as you use more and more GPUs and compute to train them. Then, after a fine- tuning process, you end up releasing them, and folks use them. That sort of scaling law has leveled off some. What is now happening is the new scaling law that has been unlocked is when you, the user, actually enter a query into, you know, a Claude or Gemini or ChatGPT or whatever your model is, it will take time to think about the answer. Then, we are scaling up the compute that is involved in that thinking process. They are called thinking or reasoning models. Just in a few weeks, whatever it was, when OpenAI went from the o1 model to the o3 model, we saw this incredible increase in capability, and there is not really any sign that it is slowing down. The implication of that is that we might see even more rapid and even more astonishing advances very soon. It is pretty astonishing what these systems can accomplish even now. I mean, I am someone who doesn't know how to write a line of code, but you can go to the latest models, and I can basically design you a rudimentary computer game instantly by just telling the model what to code. For some of them, like Claude, it will actually produce some version of it right there for you. Others, you can just cut and paste it into some other application, but then it creates the application for you. On the last chart, it said we are now kind of above the level of a Ph.D. in their field, which is a pretty high bar to begin with. The acceleration ahead could be further stimulated by the fact that this applies to many fields, but also computer science. You could well see, and, in fact, you are already seeing it to some extent, and it is likely to pick up dramatically, the models themselves that are working on their own coding that are contributing to their own growth and capability. You are also, by the way, seeing models that are built around solving specific problems. I mentioned AlphaFold or models built around discovering new materials. You also have uses of AI that are guiding action in the physical world, such as self-driving cars. If you don't live in L.A. or San Francisco or Phoenix, you probably haven't ridden in a Waymo. It works incredibly well as a driverless vehicle. It surpassed the market share of Lyft, I believe, now in San Francisco. This has been made possible largely by AI or similarly Tesla's self-driving systems available in millions of cars being trained with a neural network. Also, none of this accounts for the potential of quantum computing where Google recently achieved an incredible breakthrough and the interaction of quantum and AI could lead to even more staggering results. This acceleration is happening very, very fast. My purpose in talking about it is not a call for regulation. Although, I do think some regulation is appropriate, but ultimately no regulation is going to stop this or really even dramatically slow down this progress. I mean, the example of DeepSeek and what China is working on makes that point very clearly. I do, however, think that we need to invest heavily in research around AI safety and alignment. I th